In this paper, an improved algorithm for active control of fan noise is proposed. The main idea behind this work is to reshape the frequency components of the reference signal using the information from the unwanted fan noise. In the algorithm, the reflection coefficients of linear prediction (LP) based on lattice structure are obtained from the noise. The coefficients are used to build an autoregressive (AR) model, which is applied to the reference signal to improve the correlation between the optimized reference signal and the noise. This method of spectral reshaping for reference signal (SRRS) is carried out automatically by an online adaptive process. The SRRS improves the control performance noticeably. The algorithm has a fast convergence rate and high stability with a tolerant increase in computational complexity. It is implemented to reduce the noise from a real fan on a digital signal processor (DSP) platform. © 2014 Elsevier Ltd. All rights reserved.